A New Many-Objective Optimization Approach to Association Rule Mining: The NSGA-II/DE-ARM Algorithm

Abstract

Association rule mining is a fundamental data mining technique for uncovering latent relationships among variables in large-scale datasets. However, conventional approaches rely on single-metric filtering strategies, which are insufficient for capturing the inherent multi-criteria nature of rule quality. To address this limitation, this study formulates ARM as a many-objective optimization problem and proposes a hybrid algorithm, NSGA-II/DE-ARM, that simultaneously optimizes four rule-quality measures: support, confidence, lift, and NetConf. The proposed algorithm enhances the NSGA-II framework by integrating binary differential evolution operators, an adaptive operator selection mechanism, lift-weighted tournament selection, and a constraint-domination principle combined with a dynamic minimum support threshold. Its performance was evaluated using two datasets: a SIPRI-World Bank panel dataset consisting of defense industry and macroeconomic indicators covering 46 items over the 2002-2023 period, and the UCI Mushroom benchmark dataset consisting of 118 items. Across 30 independent runs on the SIPRI-World Bank dataset, NSGA-II/DE-ARM outperformed the Apriori baseline in all four metrics (mean lift = 4.748, confidence = 0.853, support = 0.146, NetConf = 0.789), with large effect sizes (Cohen's d = 1.77-5.77, p < 0.001 in each case). On the Mushroom benchmark dataset, the proposed method also achieved substantial improvements, with Cohen's d values ranging from 0.93 to 6.16. NSGA-II/DE-ARM generated 68 Pareto-optimal rules in a representative run and achieved the highest hypervolume values on both datasets, with HV = 3.231 for SIPRI-World Bank and HV = 6.262 for Mushroom. These results suggest that NSGA-II/DE-ARM offers decision-makers a broader and more balanced multi-criteria solution set than single-metric filtering approaches.